DiscoDemo: Discovering Efficient and Diverse Robot Demonstrations for Imitation Learning

DiscoDemo. Discovering Efficient and Diverse Robot Demonstrations for Imitation Learning. By Minho Park and eight others, from KAIST and Holiday Robotics. Published on the twenty-third of October. Its code repository is worked out: about four hundred and fifty source files, with a readme and a licence.

Behavior cloning for robotic manipulation is highly sensitive to how well its training data covers the task. Collecting diverse successful demonstrations, though, remains costly. The premise of this work is that coverage comes from variety in how a task gets solved. Several successful solution modes cover a broader range of successful trajectories than collecting one narrow solution over and over again. The limit, in other words, is not only how many demonstrations there are, but how differently they solve the task.

DiscoDemo puts reinforcement learning in a specific role. It is not used as a controller to deploy. It is used as a data generator, whose output feeds downstream visuomotor imitation learning. The generator is trained to discover diverse solutions while preserving task success, so that what it hands over is many different ways of succeeding rather than many copies of one. The RL policy is judged not by how it would perform as a controller, but by the data it leaves behind, and by how much of the range of successful trajectories that data covers.

The generator is a privileged, latent-conditioned RL generator. It is trained with a parallel reverse curriculum, and with a success-gated diversity objective, so that diversity is pursued while success is kept. Once trained, it produces multiple successful closed-loop behaviors. Those behaviors are safety-filtered and then exported as a static dataset, and that dataset is what the downstream imitation learning trains on.

The evaluation covers four manipulation tasks. DiscoDemo produces broader coverage of successful states, and it improves downstream visuomotor policies both in simulation and in the real world. Averaged across tasks in simulation, downstream success rises to sixty-two percent, from forty percent for the strongest baseline.

On the real robot, averaged across tasks, success rises to seventy-eight percent, from thirty-nine percent for the strongest baseline. Both settings point the same way: broader coverage of successful states going in, and better downstream policies coming out.

Seventy-eight percent against thirty-nine on the real robot is roughly a doubling, and the change behind it is in the data: what the reinforcement learning generator chose to demonstrate. The authors read these results as evidence that diversity among successful solutions improves imitation learning for a pretrained flow-matching VLA, and that this diversity remains complementary to corrective data collection rather than a replacement for it.